microideax / FeedforwardNet

Bottom up designed feed forward network functions for CNN applications

Table of Contents

FeedforwardNet is a toolkit that simplifies the development and deployment of deep learning systems for FPGA based devices, either embedded or resource rich. The toolkit enables fast mapping pre-trained network models or applications on FPGA platforms and accelerate with pure logical hardware.

This toolkit contains a bottom up designed inference network library together with the pre-trained CAFFE model converter and network construction flow. This version of code is still under active development. Future version of well integrated toolkit with a web based GUI will be released soon.

1. caffe converter -- converting caffemodels to weight, mean, val and network parameter files.
2. codeGenerator -- generating the targeted network model in C++.
3. example -- several example caffemodels and generated projects.
4. fpga_cnn -- original accelerator template lib.
5. scripts -- scripts used to organize the generation flow.
6. stb_image -- stb image lib with non-preinstallation required.

Getting Started

1. compile caffe_converver and convert the caffemodel file:

• cd to caffe_converter.
• run ./run.sh to compile the converter codes.
• use ./caffe_converter targeted.prototxt targeted.caffemodel [targeted.binarymodel] to convert the input caffemodel. The detailed explanation and instructions are located in caffe_converter.

2. generate the targeted network with the net_config_params.txt from caffe_converter:

• navigate to codeGenerator folder.
• cp ../caffe_converter/net_config_params.txt ./
• run ./run_generator.sh and follow the instruction with correct paths,file names and data types. Make sure the test_demo folder in example/ is deleted before starting this step. The detailed explanation and instructions are located in codeGenerator.

3. test the generated network in C++:

• navigate to FeedforwardNet/example/, the newly generated network is located here in test_demo/.
• compile the C++ code with command make.
• run ./ff_test to test the correctness.

4. FPGA synthesis and implementation:

• uncomment HLS_MODE in config.h located in inference_net/.
• navigate into hls_impl/ folder.
• run ./syn.sh to start synthesis and generate the inference_net IP.

Help and Support

If you have any suggestions and questions, please contact [e-mail].

Bibtex:

@misc{feedforward2016,

author = {Yao Chen, Yang Yu, Chunrong Zhong},

title = {{FeedforwardNet: Enabling Efficient Convolutional Neural Network Application Design on Embedded FPGAs, Version 0.1}},

year = {2016},

url = {https://github.com/microideax/FeedforwardNet} }